AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

NRGPT: An Energy-based Alternative for GPT

Nima Dehmamy, Benjamin Hoover, Bishwajit Saha, Leo Kozachkov, Jean-Jacques Slotine, Dmitry Krotov

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Efficient language transformers · 5 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use NRGPT energy-based block instead of recurrent GPT blockrecurrent GPT (GPT_Rec_parallel) · OpenWebText n_embed=768validation loss: 3.454 -> 3.467No clear effect
use NRGPT instead of 12-layer GPTGPT (124M) · OpenWebText n_embed=768validation loss: 2.921 -> 3.467Hurt
use NRGPT instead of GPT for generationGPT · OpenWebText generationdistinct-1: 0.619 -> 0.674 (distinct-1); perplexity 75 -> 104Mixed
use NRGPT instead of GPT_Rec_parallel on ListOpsrecurrent GPT · ListOpsparameters at learning transition: 2.3 -> 2.4 (FF1) / 2.98 (FF2W), x10^4No clear effect
use NRGPT instead of GPT on Shakespeare at large sizesGPT · Shakespeareoverfitting / validation lossImproved
widen FF1 hidden dimension from 4D to 8D (ablation)NRGPT_H_FF1 · ListOpscompetitiveness with Rec-GPTImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy